Mesh Convergence

Mesh convergence verifies that a numerical result is not primarily an artefact of the chosen element size, cell size, or interpolation order. In a finite element model, for example, a coarse mesh may underpredict peak stress, over-stiffen a structure, smear a thermal gradient, or miss a local contact effect. Refining the mesh should move the selected result toward a stable value.

The target of the study is not the entire model in the abstract. It is a quantity of interest: maximum displacement at a bracket tip, reaction force at a support, pressure drop through a duct, stress intensity near a crack, heat flux through an interface, or natural frequency of a mode. Different quantities can converge at different rates, so a model may be adequate for stiffness but inadequate for local fatigue assessment.

Refinement strategies

The common approach is an h-refinement study: the mesh is made progressively smaller while element type, solver settings, geometry, loads, and boundary conditions remain controlled. A p-refinement study increases polynomial order instead of reducing element size. Some solvers use adaptive refinement, where error indicators concentrate new elements in regions with high gradients or residuals. In each case, the engineer plots the quantity of interest against mesh density and checks whether the remaining change is small enough for the decision being made.

Good convergence studies keep non-mesh variables fixed. Changing contact stiffness, turbulence model, time step, material law, or load application while refining the mesh makes the result impossible to interpret. Mesh quality also matters: skewed elements, high aspect ratios, poor transition zones, or badly aligned cells can degrade convergence even when nominal element count increases.

Singularities and local peaks

Some results do not converge to a finite value because the model contains a mathematical singularity. Sharp re-entrant corners, point loads, perfectly fixed edges, idealized bonded contacts, and abrupt material changes can make local stress grow indefinitely as the mesh is refined. In those cases the convergence study should shift to an averaged stress, structural stress, energy release quantity, displacement, reaction load, or another physically meaningful metric.

Mesh convergence is a verification activity, not full validation. A converged model can still be wrong if the boundary conditions are unrealistic, the constitutive model is inappropriate, the geometry omits a critical feature, or the operating load is poorly estimated. It answers “is the discrete solution stable enough?” rather than “is the model true to the physical system?”

Common mistakes

A weak convergence claim often reports only total element count. Better documentation states element type, refinement pattern, mesh-quality checks, solver tolerances, the plotted quantity of interest, and the acceptance threshold. Another common error is using a visually smooth contour plot as evidence of convergence. Smoothness can be produced by post-processing interpolation even when the underlying numerical error is still large.